This is a minor release implementing new functionality, and including bug fixes, updates to documentation, argument checking and test coverage.
Added the rescale_probs argument to the setup_trial() family of
functions, allowing automatic rescaling of fixed allocation probabilities
and or minimum/maximum allocation probability limits when arms are dropped
in simulations of trial designs with >2 arms.
The extract_results() function now also returns errors for each simulation
(in addition to squared errors) and the check_performance(),
plot_convergence(), and summary() functions (including their print()
methods) now calculate and present median absolute errors (in addition to
root mean squared errors).
The plot_metrics_ecdf() function now supports plotting errors (raw,
squared, and absolute), and now takes the necessary additional arguments
passed to extract_results() used for arm selection in simulated trials not
stopped for superiority.
Added the update_saved_calibration() function to update calibrated trial
objects (including embedded trial specifications and results) saved by the
calibrate_trial() function using previous versions of the package.
Rewritten README and 'Overview' vignette to better reflect the typical usage and workflow.
The setup_trial() family of functions now stops with an error if less than
two arms are provided.
The setup_trial() family of functions now stops with an error if
control_prob_fixed is "match" and fixed_probs is provided for the
common control arm.
Improved error message when true_ys-argument is missing in
setup_trial_binom() or when true_ys- or sds-argument is missing in
setup_trial_norm().
Changed the number of rows used in plot_convergence() and plot_status()
if the total number of plots is <= 3 and nrow and ncol are NULL.
Fixed a bug in extract_results() (and thus all functions relying on it),
causing arm selection in inconclusive trial simulations to error when
stopped for practical equivalence and more simulated patients were
randomised than included in the last analysis.
Improved test coverage.
Minor edits and clarification to package documentation.
Added references to two open access articles (with code) of simulation
studies using adaptr to assess the performance of adaptive clinical trials
according to different follow-up/data collection lags
(https://doi.org/10.1002/pst.2342) and different sceptical priors
(https://doi.org/10.1002/pst.2387)
This is a patch release with bug fixes and documentation updates.
Fixed a bug in check_performance() that caused the proportion of
conclusive trial simulations (prob_conclusive) to be calculated
incorrectly when restricted to simulations ending in superiority or with a
selected arm according to the selection strategy used in restrict. This
bug also affected the summary() method for multiple simulations (as this
relies on check_performance()).
Fixed a bug in plot_convergence() that caused arm selection probabilities
to be incorrectly calculated and plotted (this bug did not affect any of the
other functions for calculating and summarising simulation results).
Corrections to plot_convergence() and summary() method documentation for
arm selection probability extraction.
Fixed inconsistency between argument names and documentation in the internal
%f|% function (renamed arguments for consistency with the internal %||%
function).
This is a patch release triggered by a CRAN request to fix a failing test that also includes minor documentation updates.
Fixed a single test that failed on CRAN due to an update in the testthat
dependency waldo.
Fixed erroneous duplicated text in README and thus also on GitHub and the package website.
Minor edits/clarifications in the documentation including function documentation, README, and vignettes.
This release implements new functionality (most importantly trial calibration), improved parallelism, a single important bug fix, and multiple minor fixes, changes, and improvements.
Added the calibrate_trial() function, which can be used to calibrate a
trial specification to obtain (approximately) a desired value for a certain
performance characteristic. Typically, this will be used to calibrate trial
specifications to control the overall Bayesian type-1 error rates in a
scenario with no between-arm differences, but the function is extensible and
may be used to calibrate trial specifications to other performance metrics.
The function uses a quite efficient Gaussian process-based Bayesian
optimisation algorithm, based in part on code by Robert Gramacy (Surrogates
chapter 5, see: https://bookdown.org/rbg/surrogates/chap5.html), with
permission.
More and better parallelism. The functions extract_results(),
check_performance(), plot_convergence(), plot_history(), and the
summary() and print() methods for trial_results objects may now be run
in parallel via the cores argument or as described below. Please note, some
other functions have not been parallelised, as they were already fast and
the time it took to copy data to the clusters meant that the parallel
versions of those functions were actually slower than the original ones,
even when run on results from 10-100K simulations.
The setup_cluster() function has been added and can now be used to setup
and use the same parallel cluster throughout a session, avoiding the
overhead of setting up and stopping new clusters each time. The default
value for the cores argument in all functions is now NULL; if an actual
value is supplied, it will always be used to initiate a new, temporary
cluster of that size, but if left at NULL the defaults defined by
setup_cluster() will be used (if any), otherwise the "mc.cores" global
option will be used (for new, temporary clusters of that size) if specified
by options(mc.cores = <number>), and otherwise 1.
Finally, adaptr now always uses parallel (not forked) clusters as is the
default in parallel and which works on all operating systems.
Better (safer, more correct) random number generation. Previously, random
number generation was managed in an ad-hoc fashion to produce similar
results sequentially and in parallel; while the influence of this was
minimal, the package now uses the "L'Ecuyer-CMRG" random number generator
(see base::RNGkind()) and appropriately manages random number streams
across parallel workers, also when run sequentially, ensuring identical
results regardless of the use of parallelism or not.
Important: Due to this change, simulation results from run_trials()
and bootstrapped uncertainty measures from check_performance() will not be
identical to those generated with previous versions of the package. In
addition, individual trial_result objects in a trial_results object
returned from run_trials() no longer contain any individual seed values,
but instead NULL.
Added the plot_metrics_ecdf() function, which plots empirical cumulative
distribution functions of numerical performance metrics across multiple
trial simulations.
Added the check_remaining_arms() function, which summarises all
combinations of remaining arms across multiple simulations.
Fixed a bug in extract_results() (and thus also in functionality relying on
it: check_performance(), plot_convergence(), and the summary()
method for multiple simulated trials) that caused incorrect total event
counts and event rates being calculated for trial specification with
follow-up/outcome-data lag (the total event count from the last adaptive
analysis was incorrectly used, and for ratios this was divided by the total
number of patients randomised). This has been fixed and the documentation of
the relevant functions has been updated to clarify the behaviour further.
This bug did not affect results for simulations without follow-up/outcome-
data lag.
Values for inferiority must now be less than 1 / number of arms if no
common control group is used, and the setup_trial() family of functions
now throws an error if this is not the case. Larger values are invalid and
could lead to simultaneous dropping of all arms, which caused run_trial()
to crash.
The print() method for results from check_performance() did not respect
the digits argument; this has been fixed.
Now includes min/max values when summarising numerical performance metrics
in check_performance() and summary(), and these may be plotted using
plot_convergence() as well.
The setup_trial() functions now accepts equivalence_prob and
futility_prob thresholds of 1. As run_trial() only stops or drops arms
for equivalence/futility if the probabilities exceed the current threshold,
values of 1 makes stopping impossible. These values, however, may be used
in a sequence of thresholds to effectively prevent early stopping for
equivalence/futility but allowing it later.
When overwrite is TRUE in run_trials(), the previous object will be
overwritten, even if the previous object used a different trial
specification.
Various minor updates, corrections, clarifications, and structural changes to package documentation (including package description and website).
Changed size to linewidth in the examples in plot_status() and
plot_history() when describing further arguments that may be passed on to
ggplot2 due to deprecation/change of aesthetic names in ggplot2 3.4.0.
Documentation for plot_convergence(), plot_status(), and
plot_history() now prints all plots when rendering documentation if
ggplot2 is installed (to include all example plots on the website).
The setup_trial() functions no longer prints a message informing that
there is no single best arm.
Various minor changes to print() methods (including changed number of
digits for stopping rule probability thresholds).
The setup_trial() family of functions now restores the global random seed
after being run as the outcome generator/draws generator functions are
called during validation, involving random number generation. While this was
always documented, it seems preferable that to restore the global random
seed during trial setup when functions are only validated.
Always explicitly uses inherits = FALSE in calls to base::get(),
base::exists(), and base::assign() to ensure that .Random.seed is only
checked/used/assigned in the global environment. It is very unlikely that
this would ever cause errors if not done, but this serves as an extra
safety.
This is a minor release implementing new functionality, updating documentation, and fixing multiple minor issues, mostly in validation of supplied arguments.
Simulate follow-up (and data collection) lag: added option to have
different numbers of simulated patients with outcome data available
compared to the total number of simulated patients randomised at
each adaptive analysis (randomised_at_looks argument in
setup_trial() family of functions). Defaults to same behaviour as
previously (i.e., assuming that outcome data are immediately
available following randomisation). As a consequence, run_trial()
now always conducts a final analysis after the last adaptive
analysis (including both final posterior and 'raw' estimates),
including the outcome data of all patients randomised to all arms,
regardless of how many had outcome data available at the last
conducted adaptive analysis. Both sets of results are saved and
printed for individual simulations; extract_results(), the
summary() and the print() methods for multiple simulations have
gained the additional argument final_ests that controls whether
the results from this final analysis or from the last relevant
adaptive analysis including each arm are used when calculating some
performance metrics (defaults are set to ensure backwards
compatibility and otherwise use the final estimates in situations where
not all patients are included in the final adaptive analysis). An
example has been added to the Basic examples vignette illustrating
the use of this argument.
Updated plot_history() and plot_status() to add the possibility
to plot different metrics according to the number of patients
randomised as specified by the the new randomised_at_looks
argument to the setup_trial() functions as described above.
Added the update_saved_trials() function, which 'updates' multiple
trial simulation objects saved by run_trials() using previous
versions of adaptr. This reformats the objects to work with the
updated functions. Not all values can be added to previously saved
simulation results without re-running; these values will be replaced
with NAs, and - if used - may lead to printing or plotting of
missing values. However, the function allows re-use of the same data
from previous simulations without having to re-run them (mostly
relevant for time-consuming simulations). Important: please
notice that other objects (i.e., objects returned from the
setup_trial() family of functions and single simulations returned
by run_trial()) may create problems or errors with some functions
if created by previous versions of the package and manually
reloaded; these objects will have to be updated by re-running the
code using the newest version of the package. Similarly, manually
reloaded results from run_trials() that are not updated using this
function may cause errors/problems when used.
Added the check_performance() function (and a corresponding
print() method) which calculates performance metrics and can be
used to calculate uncertainty measures using non-parametric
bootstrapping. This function is now used internally by the
summary() method for multiple trial objects.
Added the plot_convergence() function which plots performance
metrics according to the number of simulations conducted for
multiple simulated trials (possibly after splitting the simulations
into batches), used to assess stability of performance metrics.
Added the possibility to define different probability thresholds for
different adaptive analyses to the setup_trials() family of
functions (for inferiority, superiority, equivalence, and futility
probability thresholds), with according updates in run_trial() and
the print() method for trial specifications.
Updated plot_status(); multiple arms may now simultaneously be
plotted by specifying more than one valid arm or NA (which lead to
statuses for all arms being plotted) in the arm argument. In
addition, arm name(s) are now always included on the plots.
Added reference to open access article describing key methodological considerations in adaptive clinical trials using adaptive stopping, arm dropping, and randomisation to the package documentation (https://doi.org/10.1016/j.jclinepi.2022.11.002).
The proportion of conclusive trials when restricting the trials
summarised (in extract_results()) may now be calculated by the
summary() method for multiple trial simulations and by the new
check_performance() function, even if this measure may be
difficult to interpret when the trials summarised is restricted.
Minor fixes, updates, and added clarification to the documentation in multiple places, including in vignettes, which have also been updated to illustrate some of the new functions added.
Minor fix to print() method for individual trial results, which did
not correctly print additional information about trials.
Fixed a bug where the same number of patients included could be used
for subsequent data_looks in the setup_trial() family of
functions; this now produces an error.
Added internal vapply_lgl() helper function; the internal vapply()
helper functions are now used consistently to simplify the code.
Added multiple internal (non-exported) helper functions to simplify
code throughout the package: stop0(), warning0(), %f|%, and
summarise_num().
Added names = FALSE argument to quantile() calls in the
summary() method for trial_results objects to avoid unnecessary
naming of some components if they are subsequently extracted from
the returned object.
Ideal design percentages may be calculated as NaN, Inf or -Inf
in scenarios with no differences; these are now all converted to
NA before being returned by the various functions.
Minor edits/clarifications to several errors/warnings/messages.
Minor fix to internal verify_int() function; when supplied with,
e.g., a character vector, execution was stopped with an error
instead of returning FALSE, as needed to print the proper error
messages after checks.
Minor fix to plot_status(), where the upper area (representing
trials/arms still recruiting) was sometimes erroneously not plotted
due to a floating point issue where the summed proportions could
sometimes slightly exceed 1.
Added additional tests to test increase coverage of existing and new functions.
Minor fix in internal reallocate_probs() function, when "match"-ing
control arm allocation to the highest probability in a non-control arm and
if all probabilities were initially 0, the returned vector lacked names,
which have now been added.
Minor fixes to internal validate_trial() function in order to: not
give an error when multiple values were supplied to the
control_prob_fixed argument; and to give the correct error when
multiple values were provided to equivalence_diff or futility_diff;
and to give an error when NA was supplied to futility_only_first;
and to add some tolerance to the checks of data_looks and
randomised_at_looks to avoid errors due to floating point imprecision
when specified using multiplication or similar; and correct errors if
decimal numbers for patient count arguments were supplied; and additional
minor updates to errors/messages.
This is a patch release triggered by a CRAN request for updates.
Minor formatting changes to the adaptr-package help page to comply
with CRAN request to only use HTML5 (as used by R >=4.2.0).
Minor bug fixes in print() methods for trial specifications and
summaries of multiple trial results.
Minor updates in messages in setup_trial().
Minor release:
Updates to the run_trials() function to allow exporting objects to
clusters when running simulations on multiple cores.
Updates to internal function verify_int() due to updates in R >=
4.2.0, to avoid incorrect error messages in future versions due to
changed behaviour with the && function when used with arguments
with length > 1
(https://stat.ethz.ch/pipermail/r-announce/2022/000683.html).
Minor documentation edits and updated citation info (reference to software paper published in Journal of Open Source Software, https://doi.org/10.21105/joss.04284).